Europe's data centre constraints may become its AI infrastructure advantage

Europe is not winning the race to build artificial intelligence data centres. The United States operates roughly 5,400 facilities; most European countries have two to three hundred. McKinsey estimates the world must double or triple the entire data centre stock built over the past forty years, a $7 trillion undertaking by 2030. Yet the very constraints that make Europe slow, fragmented grids, expensive power, strict planning, are starting to look like a form of future-proofing.

The energy bottleneck splits the continent

Electricity access is the defining chokehold. The Nordics and Spain have seen appetite surge because they possess surplus generation, hydropower in the north, solar and wind in the south. Germany, the United Kingdom, Ireland and the Netherlands sit on the other side of the ledger. "Basically either we just don't have the grid capacity right now or we've got such a shortage in the system that there's effectively a moratorium for the foreseeable future," said Jags Walia, head of global listed infrastructure at Van Lanschot Kempen. Italy occupies a middle ground: its average grid connection time is three years, against a European mean of four, according to energy think tank Ember.

The International Energy Agency projects data centre electricity consumption will more than double to 1,000 terawatt-hours in 2026, up from 460 TWh in 2022, driven largely by AI workloads. Power is the single largest cost component in a modern facility. The UK now has the highest industrial energy costs in Europe, roughly 75 per cent above pre-invasion levels. That differential shapes where capital flows.

Grid queues and the speculator problem

Congestion has created a secondary market in grid connections. "You get a lot of speculators in the queue, and those speculators make it more difficult because they have no intention of building data centres. They just want the power, perhaps, to flip it somebody else," said Kevin Restivo, European data centre research lead at CBRE. The UK historically allocated connections on a first-come-first-served basis, ignoring project significance. It is now transitioning to a "first ready, first connected" model that lets finished projects jump the queue, a reform designed to flush out speculators and force developers to move faster once they have a slot.

Last year the UK government designated data centres Critical National Infrastructure, allowing central authorities to overrule local planning refusals. The move signals how seriously Whitehall treats the asset class, but it also underscores the political sensitivity of land use in a crowded island.

Diversification beyond the FLAP-D core

The traditional FLAP-D markets, Frankfurt, London, Amsterdam, Paris, Dublin, are saturating. Investment is migrating to secondary regions where power and fibre are plentiful. Walia argues the quickest route around grid constraints is not to wait for new connections but to ask: "Where do I currently have good grid connection to an industry in decline?" Brownfield sites, former aluminium smelters, chemical plants, automotive factories, already possess heavy grid connections and can be repurposed as tech hubs. The approach is slower than greenfield US builds, but it avoids the years-long wait for new transmission infrastructure.

Inference, not training, is Europe's lane

Consensus among investors is that Europe will not lead in the massive campuses required for frontier model training. That race is considered all but won by US hyperscalers. The continent has few foundational model developers, France's Mistral is the most prominent, but McKinsey sees 70 per cent of all AI demand coming from inference, the act of running trained models. Inference workloads need high fibre density, proximity to users, and often must reside within national borders for data sovereignty reasons.

"So, actually, you are finding these areas, from our perspective, are well protected from that potential oversupply bubble that could come through," said Seb Dooley, senior fund manager at Principal Asset Management. He expects inference to sit in the same facilities as cloud, a pattern already visible at some US sites. That gives investors "quite a nice upside" without the speculative risk attached to pure-play AI builds. Inference racks run hotter, densities above 20 kilowatts per rack versus typical cloud loads, and require liquid or immersion cooling. Designing for both workloads from the start is becoming a prerequisite.

Regulation as a moat

European policy demands developers report energy and water usage and justify site selection. Spain has proposed adding socio-economic impact assessments. "Nobody asks about that in the US," Walia noted. Dooley argues the friction works in Europe's favour over the long run: facilities become integrated into local communities "rather than just being a complete blight on everyone's life that they can sometimes be." Sustainability is one area where the bloc has been "very good at innovating." The European Commission's energy efficiency and taxonomy rules effectively set a floor under build quality that US developers only meet voluntarily.

Speculative builds, erecting a shell and hunting for tenants later, are becoming "a relic of the past, for the most part," Restivo said. Developer-operators now lock anchor tenants into 10-to-15-year leases before breaking ground. That cushions obsolescence risk: if AI's technical needs shift, the tenant bears the retrofit cost. The exception is neo-cloud providers, start-ups with unproven models that sign five-to-seven-year terms. "There is a lot of skin in the game for developer-operators working with neo-clouds," Restivo added, though some debt financiers are growing comfortable with the shorter duration.

Sovereign AI: the underestimated driver

Governments increasingly treat domestic inference capacity as a strategic asset. Jim Wright, manager of the Premier Miton Global Infrastructure Income Fund, called sovereign AI an "underestimated" driver of the European build-out. If every major economy requires inference within its borders, the addressable market for European facilities expands regardless of whether the continent produces the underlying models. That demand is structural, not cyclical, and it favours the constrained, regulated environments where Europe already operates.